Netflix Contest
So I just read that Netflix is running a contest to try and improve their recommendation system. A prize of $1 million to the person who can improve the accuracy of their existing system by 10%. I'm assuming that by accuracy, they mean successful recommendations (that is, if the system recommends a movie to you, and you rent said movie, the recommendation is successful).
Not really a proposal for this contest, but just a general thought about online recommendation systems in general.
'Satisfaction Garunteed' is a common phrase in today's retail experience. It generally doesn't apply to digital goods, though, for fear that customers will simply copy the data and return the product.
With Netflix's existing system, there must be some recommendations they make, though, for which they have a high level of confidence that the customer will like. Others, certainly, are more 'risky'.
Why not simply broadcast their confidence heuristic with each recommendation and find a 'sweet spot' (say, 65% confidence) where they take a chance. Give these 'sweet spot' recomendations to the customer for free (that is, they don't count against the total number of reservations allowed, or some other incentive system is used to encourage the customer to try).
So what is the sweet spot then? Why bother to give anything away for free (or offer any sort of incentive to renting the title in question)? Well, the title is obviously _sort of_ in the user's interest profile, but not exactly (other wise it would be a higher confidence heuristic). If they rent it and like it, though, the additional information that could be gleamed from the rental to add to the system's understanding of their tastes could be worth the cost of the incentive.
In particular I wonder why Amazon doesn't adopt something like this. 10 day unconditional money back garuntee for non-digital goods based off of recommendation heuristics. "We're so sure in our system's ability to suggest the next book you read, we'll send it to you to try for free for 10 days".
Works for Time Life music. Why not amazon when the 'we're so sure' part actually means something.
Not really a proposal for this contest, but just a general thought about online recommendation systems in general.
'Satisfaction Garunteed' is a common phrase in today's retail experience. It generally doesn't apply to digital goods, though, for fear that customers will simply copy the data and return the product.
With Netflix's existing system, there must be some recommendations they make, though, for which they have a high level of confidence that the customer will like. Others, certainly, are more 'risky'.
Why not simply broadcast their confidence heuristic with each recommendation and find a 'sweet spot' (say, 65% confidence) where they take a chance. Give these 'sweet spot' recomendations to the customer for free (that is, they don't count against the total number of reservations allowed, or some other incentive system is used to encourage the customer to try).
So what is the sweet spot then? Why bother to give anything away for free (or offer any sort of incentive to renting the title in question)? Well, the title is obviously _sort of_ in the user's interest profile, but not exactly (other wise it would be a higher confidence heuristic). If they rent it and like it, though, the additional information that could be gleamed from the rental to add to the system's understanding of their tastes could be worth the cost of the incentive.
In particular I wonder why Amazon doesn't adopt something like this. 10 day unconditional money back garuntee for non-digital goods based off of recommendation heuristics. "We're so sure in our system's ability to suggest the next book you read, we'll send it to you to try for free for 10 days".
Works for Time Life music. Why not amazon when the 'we're so sure' part actually means something.
posted by Ben Mattes at 4:18 PM
0 Comments:
Post a Comment
<< Home